vLLM test reports 13% higher Step5 throughput with PrimTS on Blackwell

The merged, opt-in FlashInfer path reported 147.04 tokens per second at roughly 96K input and concurrency 32, and 5.438-second time to first token for a separate 160K batch-one case on four GB300s.

By · Published

Four large black pipes run from an industrial building across a gravel bed, with ventilation units and a door marked “Rw” under a sunset sky.

The merged, opt-in FlashInfer path reported 147.04 tokens per second at roughly 96K input and concurrency 32, and 5.438-second time to first token for a separate 160K batch-one case on four GB300s.